OBIA Spectral Gradient Tutorial

One of the challenges in getting started with object-based image analysis (OBIA) is the complexity inherent in high-resolution remotely sensed data sets.  Inspired by one of Bruce Gorham's OBIA workshops we started developing some examples using pseudo data.  I recently posted a tutorial rule set to the eCognition Community site that I am calling "Spectral Gradient."  This post provides an overview of that rule set.  Please note that the rule set is designed to be a self contained tutorial in that the description of each process step is present within the rule set.  To access the documentation click on the appropriate parent process then scroll to the lower right of the Process Tree window.  The show/hide comment icon will appear, click on it and the documentation will be displayed in a sub-window.

The tutorial starts with this image. 

The objective is to classify "Red Circles."  Red Circles must meet the following criteria:
- Have a pixel that is at least 90% pure red
- Not contain any pixels that are pure white
- Have the geometric attributes consistent with a circle.

In the image above, the circle on the right along with the rounded rectangle and the triangle have pixels that are at least 90% pure red.  The circle in the top left does not.


A quadtree segmentation is used to crate image objects.  While it is not the most robust segmentation algorithm, it is fast.

Pure red objects are classified using the "ratio of red" feature.  This feature measure the amount of that the red band contributes to the total brightness of the object.  Setting a ratio of red threshold of 0.9 is equivalent to saying "classify objects that are 90% or more red."

Red objects are grown into other objects providing the brightness of those objects is less than or equal to 254 (e.g. not white).  An infinite loop is used, meaning that the grow region algorithm will run until the criteria can no longer be fulfilled. 

The merge region algorithm is first used to merge objects that are classified as Red then those objects that are unclassified.  This is a necessary prerequisite if geometric attributes are to be used to classify circles.

Circles have low asymmetry and roundness values, thus both criteria are used to evaluate Red objects as circles.

The end result is that we have successfully extracted the only circle containing "pure red" pixels.